sync : ggml #3962
danbev
merged
102 commits into
master
from
sync-ggml-26-07-30
hexagon: tiling, tracing and optimizations for unary ops (llama/25474)
872f31be
ggml : process data in smaller chunks in CUDA ggml_top_k() and ggml_a…
5635f4b5
opencl: cluster-parallel decode FA for Adreno (llama/25473)
f865ccc8
ggml-et: Initial ET backend (llama/24179)
9d65d9fb
hexagon: improve ARGSORT performance for small tensors (llama/25512)
e1993f4a
opencl: add int8 dp4 dense and MoE prefill optimization for Adreno GP…
ea808537
Vulkan: route large matmuls to medium tile on Adreno (llama/24877)
4bc66b57
ggml : add GGML_OP_LIGHTNING_INDEXER that implements DeepSeek V3.2/V4…
19c96ad1
cuda: Don't crash when querying memory on device with no free memory.…
f00dc30f
gguf : reject empty metadata keys (llama/24917)
dbe941f2
sycl: add Q2_K to DMMV reorder path (llama/25064)
56e02296
sycl: add fused top-k MoE (llama/25217)
6b3f1226
gguf : add tensor shape accessor (llama/24405)
56bccf60
vulkan: Use native e2m1 and e4m3 conversions for mxfp4/nvfp4 (llama/2…
9a18c7c0
CUDA: refactor MMQ kernel configuration (llama/24127)
dac3141a
metal : add Q2_0 support (llama/25419)
f5f0b641
sycl: set fattn_vec_nthreads to 256 for Battlemage (llama/25205)
e4567a66
kleidiai : add SME2 f32 kernel (llama/24414)
7f752292
ggml: uniformize im2col dst_type for all conv ops (llama/23660)
c766b6bf
ggml : add a set of functions for checking contiguity of inner tensor…
387aa894
vulkan/cpu: Support f16 as SET_ROWS src. (llama/25432)
d4c391bb
opencl: fix a dp4a bug for devices where cl_khr_integer_dot_product i…
0ffd4992
hexagon: fix hmx-queue signal enum-narrowing problem (llama/25677)
d3cdee12
opencl: avoid the vec path in GEMV for unaligned row stride (llama/25…
d44b96cc
opencl: handle OOB write in noshuffle GEMV kernels (odd ne01) (llama/…
ff6847f4
opencl: do not use `clCreateBufferWithProperties` when targeting CL 2…
a46de2e6
Flash Attention with XMX engine via oneDNN (llama/25222)
4ff381d6
sycl: Increase minimum buffer size for USM system allocations (llama/…
af1fe65c
sycl : implement xielu op (llama/25550)
15b52ca5
sycl : support kernel type fp16 for conv2d_dw (llama/25653)
b446531f
sycl : fix get_rows Q2_K, Q4_K, Q5_K (llama/25656)
af80afa4
ggml: add f16 out_prod support for CPU and out_prod op for Vulkan (ll…
aadab83f
metal: fuse snake activation (mul, sin, sqr, mul, add) (llama/25459)
e376ba3a
cuda : relax tensor contiguity requirements for quantized concat (lla…
8096ba6a
CUDA: tighter MMQ src1 buffer size for native fp4 (llama/25613)
f9d97c34
opencl: fix two issues on flash attention for Adreno a7x (llama/25697)
c4a15ec1
cuda : CUDA GGML_OP_LIGHTNING_INDEXER implementation (generic vector …
7ea23ebc
opencl: exclude some moe kernels on Adreno a7x (llama/25698)
b8c8175f
cuda: extract Q1_0 elements via __byte_perm (llama/25628)
797cccb6
opencl: disable FA and MoE weights repack to work around compiler iss…
ea8708a5
CUDA: dedup MoE gate/up activation quantization (llama/25441)
cf26eac4
ggml-cuda : restore prop.integrated on HIP builds (llama/24233)
73c71d17
Enable CUDA graphs on volta+turing (llama/25749)
aa3743a5
CUDA: Support CUDA Virtual Devices (llama/25228)
b4c0492d
vulkan: when using transfer queue for async copies, sync on event_wai…
913750f4
kleidiai: Add SME vs SME2 distinction in kernel dispatch (llama/25478)
b90a4f48
hexagon: L2 cache handling rework (dirty bit tracking with lazy flush…
1a4b5d11
DeepseekV4: Add fused hyper-connection ops (llama/25585)
070e5726
docs: added a note about using OpenCl with Adreno 810 (llama/25786)
5e5148f9
opencl: loads quants as uint for q4_K and q5_K flat mv (optimization …
4dbbaeff
opencl: add ABS op (llama/25115)
4db5bc44
sycl: fix row calculation when K_QUANTS_PER_ITERATION is 1 (llama/25690)
ecbac664
vulkan: Support Q2_0 (llama/25430)
ca31169a
ggml-blas: default hadamard mul_mat to cpu routine (llama/25710)
6c8bac6e
ggml : bump version to 0.17.0 (ggml/1568)
9958d501
opencl: transpose q4_K noshuffle scales for coalesced reads (llama/25…
7ff3d477
opencl: read/write MoE dp4a activation tiles to local memory as 128-b…
1e7873f5
opencl: load and use `kernel_gemm_moe_q6_k_f32_ns` from bin kernel li…
a23301f1
opencl: Support broadcast for Adreno MUL_MAT and honor `view_offs` fo…
6963f93f
hexagon: add CLAMP op (llama/25934)
52697549
CUDA: vectorize same-type get_rows with int4 copy (llama/25929)
9747391d
ggml-openvino: Add GGML_BACKEND_DL_IMPL invocation for OpenVINO backe…
2fd0abdc
vulkan: Refactor vk_queue to use per-instance mutexes and unique hand…
338e07dc
kleidiai : warn once when a weight type has no KleidiAI kernel (llama…
812ad391
cuda: add sqrt_softplus in topk-moe for dsv4 (llama/25896)
7c9d38e5
hexagon: check tensor type when reusing descriptors (llama/25968)
9145f479
cuda: GET_ROWS quants (llama/25962)
f74102e4
webgpu : add CONV_2D_DW (depthwise conv2d) kernel (llama/25847)
fcb2236a
ggml: enable PowerPC backend variants on AIX (llama/25983)
4d546f4a
hexagon: activation ops update (llama/25974)
613a693d
CUDA: Improve NVFP4 W4A4 activation quantization (llama/25730)
818f9e97
metal : add f16 type support to leaky relu (llama/25981)
57168708
CUDA: fix external compilation of q1_0 MMQ (llama/25778)
49d48deb
hexagon: fix Windows crash when op_poll is enabled (llama/26029)
47a3239d
hexagon: further improved pipeline of the core bits (L2, DMA, MM, FA)…
ddc63beb
opencl: do not treat NULL-mask flash attention as causal (llama/25771)
ce735813
opencl: cache compiled cl_program binaries on disk (llama/26050)
526d0ed9
HIP: remove rocWMMA FlashAttention (llama/26046)
9dbbfa47
Update ggml/src/gguf.cpp : Defined virtual keyword for destructor of …
6d4dc97c
hexagon: partial im2col support (llama/26007)
a798a7e3
opencl: fix fused RMS norm mul view offset (llama/26085)
2e5b78c1
ggml-cpu: Enable BF16 tiled gemm optimization on PowerPC (llama/26068)
98a532e1
ggml : adjust logic for offloading ops to weight's backend (llama/25832)
2fa1559f
sycl(build): parallelize ocloc invocations (llama/25903)
7ebe1c2b
Disable -ffast-math on HIP (llama/25495)
0c1bd5dc
ggml-metal: FWHT kernel for metal backend (llama/25924)
00009c89
sycl: fix use-after-return of the SDPA scale in the oneDNN flash-atte…
e43559f5
ggml-cuda: add chunked SSD matmul for Mamba-2 prefill acceleration (l…
1e4d28c3
vulkan: add iq4_nl support back to FA (llama/24585)
cb8a0ad8
ggml : set output of view src (llama/25729)
98dc1760
opencl: skip the Adreno KQ/KQV image kernels for multi-stream batches…
3583bc38
ggml-webgpu: Fix some binding alias issues to support all archs, fix …
2709c63e
add rdna3.5, and 3 to mmq configs so they can be tuned independently.…
3ab356fe
RPC: add tensor_memset (llama/25912)
58e109a9
sycl: contiguous fast path + 32-bit index math for unary elementwise …
e103c721
ggml-cuda : disable MMQ on devices with less than 48 KiB shared memor…
a816f406
ggml : Fix issue with kleidiai ci and stringop overflow warning (llam…
c2b2ecb6
metal: fix memory unwire if model is freed without any GPU operations…
2b8f0f18
CUDA: add Q2_0 support (llama/25707)
6366ca02
ggml : bump version to 0.18.0 (ggml/1576)
6f6960d6
sync : ggml
2f1f1a86
danbev
approved these changes
on 2026-07-30
parakeet : update parakeet test models generation
0aec3b24
Login to write a write a comment.
Login via GitHub